Application of lightweight neural networks for real-time image enhancement on mobile devices

Abstract

Mobile devices have transformed markedly over the past decade – their growing computational power has made it both possible and practical to handle image enhancement directly on-device, without routing data through remote servers. This article focuses on lightweight neural networks as a tool for real-time image enhancement under the hardware constraints of mobile platforms. The analysis centers on architectural approaches aimed at reducing parameter counts, lowering computational load, and minimizing energy consumption while maintaining acceptable output image quality. The tasks under consideration include sharpness enhancement, noise suppression, brightness and contrast correction, and general visual quality improvement. Particular attention is given to model optimization techniques – quantization, pruning, and dimensionality reduction – alongside the practical challenges of deploying such models in mobile software environments. Compact convolutional architectures are examined not in isolation but in relation to concrete application scenarios: mobile photography, augmented reality, telemedicine, and intelligent visual systems. The results indicate that lightweight deep learning models can deliver real-time image processing at a reasonable balance between output quality, processing speed, and device resource consumption. Further development of specialized compact architectures tailored to the demands of mobile platforms and local image processing scenarios is identified as a promising direction for both research and applied work.